Using deep learning to assess temporal changes of suspended particles in the deep sea

Abstract

While suspended particles play many important roles in the marine environment, their concentrations are very small in the deep sea, making observation difficult with existing methods: water sampling, optical sensors, and special imaging systems. Methods are needed to fill the lack of environmental baseline data in the deep sea, ones that are inexpensive, quick, and intuitive. In this study we applied object detection using deep learning to evaluate the variability of suspended particle abundance from images taken by a common stationary camera, “Edokko Mark 1”. Images were taken in a deep-sea seamount in the Northwest Pacific Ocean for approximately one month. Using the particles in images as training data, an object detection algorithm YOLOv5 was used to construct a suspended particle detection model. The resulting model successfully detected particles in the image with high accuracy (AP50 > 85% and F1 Score > 82%). Similarly high accuracy for a site not used for model training suggests that model detection accuracy was not dependent on one specific shooting condition. During the observation period, the world’s first cobalt-rich ferromanganese crusts excavation test was conducted, providing an ideal situation to test this model’s ability to measure changes in suspended particle concentrations in the deep sea. The time series showed relatively little variability in particle counts under natural conditions, but there were two turbidity events during/after the excavation, and there was a significant difference in numbers of suspended particles before and after the excavation. These results indicate that this method can be used to examine temporal variations both in small amounts of naturally occurring suspended particles and large abrupt changes such as mining impacts. A notable advantage of this method is that it allows for the possible use of existing imaging data and may be a new option for understanding temporal changes of the deep-sea environment without requiring the time and expense of acquiring new data from the deep sea.